Update app.py

app.py

@@ -2,7 +2,6 @@ import argparse
 import gradio as gr
 import os
 from PIL import Image
-import torch
 import spaces
 
 from kimi_vl.serve.frontend import reload_javascript
@@ -157,8 +156,7 @@ def predict(
     gradio_chatbot_output = to_gradio_chatbot(conversation)
 
     full_response = ""
-    with torch.no_grad():
-        for x in kimi_vl_generate(
+    for x in kimi_vl_generate(
             conversations=all_conv,
             model=model,
             processor=processor,
@@ -182,7 +180,6 @@ def predict(
             yield gradio_chatbot_output, to_gradio_history(conversation), "Generating..."
 
     logger.info("flushed result to gradio")
-    torch.cuda.empty_cache()
 
     if is_variable_assigned("x"):
         print(